AI Accelerates Deployment in Medical Technology: Four Trends to Watch in 2025

In 2024, artificial intelligence became a focus of attention in the medical device industry. Companies such as GE HealthCare, Medtronic, and Dexcom showcased new AI features, while Stryker and Quest Diagnostics expanded their AI assets through mergers and acquisitions. Meanwhile, discussions around regulation and generative AI—models trained to create new data such as images and text—dominated medical technology industry conferences.

The U.S. Food and Drug Administration (FDA) recently clarified the information needed in future submissions, as the agency has authorized more than 1,000 AI devices. However, the FDA has not yet authorized any continuously adaptive tools or those using generative AI.

Although these guidance documents should provide some clarity for medical device developers, questions remain about how regulators will handle generative AI. The Trump administration has also brought uncertainty to AI regulation. Despite the widespread attention AI has drawn in the medical technology industry, the technology still faces adoption barriers, including a lack of insurance reimbursement.

MedTech Dive spoke with several experts about AI trends to watch in 2025. Here are their predictions:

1. New AI guidance brings clarity for device developers

Lawyers say the FDA's recent guidance on AI devices should provide developers with more clarity. In December, the agency finalized guidance on Predetermined Change Control Plans (PCCPs), a new framework that allows for pre-specified modifications after a device is on the market.

Amanda Johnston, a partner at Gardner, a law firm focused on FDA matters, expects more companies to submit PCCPs and the FDA to emphasize this new approach. "I think this is partly a requirement from the FDA," Johnston said. "I do think they will push developers to adopt this framework." Johnston added that PCCPs require more upfront work from developers, but with careful planning, they can save time and costs for post-market submissions.

In January, the FDA released draft guidance outlining the information the agency expects to see in AI device submissions and when post-market monitoring may be needed. The draft also encourages developers to consider PCCPs. Megan Robertson, a lawyer at the Washington, D.C.-based law firm Epstein Becker Green, said the latest draft guidance is something developers should "keep in their pocket" and use as a checklist when preparing submissions.

Robertson expects the number of AI device submissions to increase as companies become more familiar with the FDA's approach. She added that many products in the agency's Breakthrough Devices Program involve software or AI components.

It remains unclear how President Donald Trump's new administration will handle AI. On his first day in office, Trump revoked a comprehensive executive order on AI signed by former President Joe Biden, which had required the Department of Health and Human Services (HHS) to establish an AI task force. Earlier this month, HHS released a strategic plan in response to that executive order to oversee AI in healthcare.

Trump's nominee for FDA commissioner, Martin Makary, has said little on the topic. Makary shared a JAMA article written by Scott Gottlieb last year that called on Congress to update FDA regulations for medical AI.

"You can't simply compare the first Trump administration to this one to make specific predictions," Robertson said. "But we do think this administration may take action to roll back some of the more restrictive or controversial industry actions the FDA took during the Biden administration, such as the final clinical decision support guidance."

The FDA's 2022 final guidance clarified when certain software functions should be regulated as devices. Epstein Becker Green filed a petition in 2023 on behalf of the Clinical Decision Support Coalition asking the FDA to withdraw the guidance, saying it unnecessarily increased the regulatory burden on developers. Johnston expects AI and machine learning to remain a focus for the agency under the Trump administration. She also noted that increasingly fragmented state and federal privacy laws, which could affect AI adoption, are a topic worth watching.

2. Reimbursement challenges persist

AI features can be integrated into medical devices, such as imaging equipment, or sold as standalone software platforms. For device companies, the challenge lies in how to price these features, since insurance does not cover them.

BTIG analyst Ryan Zimmerman said that currently, the Centers for Medicare & Medicaid Services (CMS) does not provide specific reimbursement for FDA-authorized AI technologies. Zimmerman added that companies must use Medicare's New Technology Add-On Payment pathway or other workarounds to obtain coverage.

Last year, a bipartisan group of senators sent a letter to CMS calling for the creation of payment pathways for algorithm-based medical services. Later in 2024, a report from a House working group found that "CMS allows limited Medicare coverage of AI technologies" as long as services meet coverage criteria.

Zimmerman said companies are marketing AI features to hospitals to speed up processes and alleviate staffing pressures. However, Brian Anderson, CEO of the nonprofit Coalition for Health AI (CHAI), said these customers are more carefully scrutinizing AI tools to determine whether they are worth the investment. After last year's "tremendous excitement" around AI, "now we're seeing some sober perspectives: If we're going to spend significant money on these things, we need to make sure we see a financial return on investment," Anderson said. "I'm hearing health systems make more of these demands of vendors."

3. More focus on foundation models and administrative tools

Most AI tools currently regulated by the FDA are concentrated in radiology, although more tools are being used in pathology, ophthalmology, and cardiology. A growing number of companies are also using large language models for administrative tasks, such as generating clinical notes. Companies including GE HealthCare are developing other types of foundation models—large-scale models that can be used for multiple purposes, such as processing MRI images, extracting information from physician notes, or analyzing electronic health record data.

Nina Kottler, chief medical officer of clinical AI at Radiology Partners, said AI solutions for radiology have "definitely changed" over the past decade. Initially, AI tools focused on detecting or triaging specific conditions, such as software that analyzes images to detect potential stroke cases. Now, more AI solutions focus on workflow, Kottler said. "The gap between volume and capacity has been widening for years," Kottler said, adding that the mismatch is "so large that it surpasses any other use case you consider."

Kottler is focused on two types of foundation models. One is language models that can summarize text into reports. These capabilities are already being sold, such as tools developed by Rad AI that generate radiology report impressions based on findings and clinical indications. These text-based models are currently not regulated as medical devices. "Whether that will be the case in the future is a question, but for now, it's excluded," Kottler said.

Kottler is also watching vision-language models that can analyze images and generate draft reports. These would fall under device regulation. Kottler added that companies began building and testing such models last year, but they have not yet received FDA authorization. Epstein Becker Green's Robertson said developers seeking to submit medical device applications using generative AI may have more resources than last year to develop regulatory strategies. However, under the new administration, how the FDA will view the risks of generative AI models remains to be seen.

---

"At the end of the day, software may not directly make the diagnosis, but it is an important part of what physicians use to arrive at a diagnosis."

Francisco Rodríguez Campos

Chief Program Officer at ECRI

---

Francisco Rodríguez Campos, chief program officer at the patient safety organization ECRI, said that while some AI tools may be used only for administrative purposes, hospitals should still scrutinize them as rigorously as other AI devices. "I've seen too many problems," Rodríguez Campos said, adding that one hospital using a note-generation tool found it did not work well after updating to the latest version. "At the end of the day, software may not directly make the diagnosis, but it is an important part of what physicians use to arrive at a diagnosis," Rodríguez Campos said. "They are affecting the delivery of healthcare."

4. Hospitals need more information to evaluate AI tools

As AI devices become more prevalent, governance issues arise, such as who is responsible for maintaining models and ensuring they work as intended. Experts say hospitals need more information before purchasing and support with performance monitoring after purchase.

Scott Lucas, vice president of device safety at ECRI, expressed concern about the combination of several factors: the hype, promise, and rapid evolution of AI tools, along with an environment in healthcare with "too many preventable events." Whether hospitals have AI solutions or can monitor and govern them "may vary greatly depending on facility resources," Lucas said.

A recent study in Health Affairs found that among hospitals using predictive models, only 61% tested them on their own data, and fewer evaluated the models for bias. The article found that hospitals that are part of large systems and have the highest operating margins were most likely to evaluate models locally. Radiology Partners' Kottler said evaluating AI models on local data is important because performance data shared by developers with the FDA does not always generalize well to different practices. "That doesn't necessarily reflect that it will perform well on my data," Kottler said.

Radiology Partners goes through a five-step evaluation before deciding to deploy an AI model. Kottler said the practice has used this process for computer vision models that analyze images as well as large language models that fill in notes.

---

"You might have a very accurate model, but if it only finds what radiologists already find, it's not actually that helpful. You're just paying again for something you've already paid for."

Nina Kottler

Chief Medical Officer of Clinical AI at Radiology Partners

---

First, Radiology Partners looks at how the model performs on its own data, ideally using a large number of cases and natural disease prevalence. It also looks at whether the model can find cases that radiologists might otherwise miss. "You might have a very accurate model, but if it only finds what radiologists already find, it's not actually that helpful," Kottler said. "You're just paying again for something you've already paid for." Kottler also looks for "wow" cases—cases that impress radiologists—as well as pitfalls or false positives, so radiologists can understand the types of errors the model might make. Then, they summarize their findings and make a decision.

Organizations like CHAI are also advocating for tools that provide more upfront information. For example, CHAI recommends the use of model cards, which Anderson describes as "nutrition labels" that provide details such as how an AI model was trained and what datasets were used. The FDA cited model cards as a transparency tool in its January draft guidance. CHAI is also building a network of assurance labs—third-party laboratories that can objectively evaluate models on populations representing health system patient groups—to enable "a more informed procurement process," Anderson said.

After hospitals begin using AI tools, they also need to monitor them over time to ensure performance does not decline. This involves working with vendors early on, the CEO added. "These health systems don't realize... how difficult and expensive this will be, and how important it is to build strong partnerships with vendors," Anderson said. "Monitoring these models is not something you can do alone; you need that partnership."